Abstract:
Classical energy minimisation paradigm based on probabilistic models of input and output data, employs an energy functional which combines the weighted prior energy of output, with a conditional energy of the input given the output. The regularising weight controls an adequacy of solutions establishing the correspondence between input and output data, to select a unique goal solution out of a multiplicity of these solutions for most frequent ill-posed problems. Such a weight is usually chosen by hand, and different choices may affect the output in unpredictable ways. This thesis proposes an alternative Two-Stage energy minimisation paradigm by treating separately both the energy terms. The First-Stage employs a very simplistic prior to obtain and store a whole set of minimisers of the ill-posed and non- regularised initial problem. Then the second stage selects an adequate solution only within the stored set of minimisers. The proposed Two-Stage paradigm needs no regularising weight, which implicitly pursues two different goals: to choose a suitable solution amongst the set of energy minimisers and include to this set not only primary but also secondary minimisers of the unregularised energy to account for unavoidable random deviations (noise) of the input data.